Fix for triangular_solve BEF executable unit test failure on ROCm.

tensorflow · Apr 28, 2022 · 97c4c10 · 97c4c10
1 parent aae8198
commit 97c4c10
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/backends/gpu/lib/kernels/blas_kernels.cc b/backends/gpu/lib/kernels/blas_kernels.cc
@@ -222,7 +222,11 @@ static Error BlasTrsmBatch(
   const void** a_array = const_cast<const void**>(b_array + batchCount);
 
   auto side_mode = wrapper::BlasSideMode::FromOpaqueValue(*sideMode);
-  int32_t a_num_elements = side_mode == CUBLAS_SIDE_LEFT ? m * m : n * n;
+  int32_t a_num_elements = 0;
+  if (platform == wrapper::Platform::CUDA)
+    a_num_elements = side_mode == CUBLAS_SIDE_LEFT ? m * m : n * n;
+  else
+    a_num_elements = side_mode == rocblas_side_left ? m * m : n * n;
   ptrdiff_t a_batch_stride_bytes = *data_type_size_bytes * a_num_elements;
   ptrdiff_t b_batch_stride_bytes = *data_type_size_bytes * m * n;
   const char* a_ptr = static_cast<const char*>(A.pointer().raw(platform));